x
Eric Michaud on the Quantization Model of Neural Scaling, Interpretability and Grokking — LessWrong